AI Blog Writing Prompts for B2B: Brief, Draft and Review

The AI blog writing prompts that work for B2B give the model your facts and tell it what to do when a fact is missing. Everything else, including tone instructions, role-play and "write like an expert", matters far less. A model asked to write about your product with nothing to go on will produce a confident article about a typical supplier, and you are not one.

Below are six complete prompts in the order you would use them, with every placeholder explained and one filled in for a fictional supplier. Copy them as they are. They work in any general-purpose chat model.

Why one big prompt fails

"Write a 1,500-word SEO blog post about stainless steel fasteners for a B2B audience" returns a fluent article with three problems. It answers no particular buyer's question. It contains numbers and capabilities the model made up because the article seemed to need them. And you cannot tell which sentences those are.

Splitting the work fixes the third problem, which is the dangerous one. An outline is quick to check. A claim list is quick to check. A finished 1,500-word draft where any sentence might be invented is not.

Google's guidance on using generative AI content makes the same point from the other side: models can produce inaccurate text, and content should be reviewed for accuracy before publishing, metadata included.

Before any prompt: the evidence packet

Collect what the article is allowed to say. Label each item so the model can cite it and you can trace it.

Label	Contents	Who supplies it
READER	Role, what they are deciding, what they already know	Sales
FACTS	Confirmed company facts: range, limits, process, documents	Owner or sales manager
SOURCES	Excerpts from standards, datasheets, your own published pages	Engineer
SCOPE	What the article covers and excludes	You
LINKS	URLs of your pages the article may link to	You
OPEN	Questions nobody has answered yet	You

OPEN is the label people leave out and the one that prevents invention. If the minimum order quantity is unknown, write "MOQ: not confirmed" there. A model told what is unknown leaves a marker. A model told nothing fills the gap.

Leave confidential drawings, customer names and credentials out of the packet. The article does not need them.

Prompt 1: Get the facts out of your own team

Use this when the packet is thin. It turns a topic into questions for the person who knows.

I am preparing a B2B article and need to interview a colleague first.

Topic: [TOPIC]
Reader: [READER]
What I already know: [FACTS]

Write 8 interview questions for [COLLEAGUE ROLE]. Each question must
ask for something only our company can know: a limit, a condition,
a process step, a common buyer mistake, or a case where we say no.
Do not ask anything a textbook could answer. Order them from most
to least important for the reader's decision.

• [TOPIC]: the buyer's question in plain words, e.g. "which stainless grade for coastal fasteners".
• [READER]: one line, e.g. "site engineer specifying fixings for a seafront walkway".
• [FACTS]: whatever you have, even two lines.
• [COLLEAGUE ROLE]: e.g. "our applications engineer" or "the sales manager who quotes these".

Send the questions by email. The answers become FACTS.

Prompt 2: Outline from the packet

Create an outline for a B2B article using only the material below.

READER: [READER]
MAIN QUESTION: [ONE QUESTION]
SCOPE: [SCOPE]
FACTS: [FACTS]
SOURCES: [SOURCES]
LINKS: [LINKS]
OPEN: [OPEN]

Give me:
1. A two-sentence opening that answers the main question directly.
2. Four to six section headings, each a statement or a question the
   reader would ask. Under each, one line on what it covers and
   which label (FACTS or SOURCES) supports it.
3. One artifact: a table, checklist or worked example, with its
   columns or items.
4. A list of anything the outline needs that is not in FACTS or
   SOURCES. Do not fill these gaps yourself.

Do not invent company capabilities, numbers, customers or quotes.

• [ONE QUESTION]: exactly one. If you have two, you have two articles.
• [SCOPE]: include the exclusions, e.g. "fasteners only; not structural design; not duplex grades".

Read item 4 first. If the list is long, the article is not ready; go back to Prompt 1. Ask for a second outline with a different structure and compare the two before choosing. The comparison often shows which sections each version padded.

Prompt 3: Draft from the approved outline

Write the article from the approved outline below, using only FACTS
and SOURCES.

OUTLINE: [APPROVED OUTLINE]
FACTS: [FACTS]
SOURCES: [SOURCES]
LINKS: [LINKS]
OPEN: [OPEN]

Rules:
- Open with the direct answer. No scene-setting introduction.
- After any sentence that states a number, a capability or a
  technical property, add the label it came from in square
  brackets, e.g. [FACTS] or [SOURCES-2].
- A property of a raw material is not proof of how our finished
  product performs. Do not make that jump.
- Keep units, conditions and words like "typically" exactly as
  they appear in the source.
- Where information is missing or two sources conflict, write
  [CHECK: what is missing] and move on. Never estimate.
- Any illustrative example must be introduced with "Example:" and
  must not name a real company.
- Plain, specific sentences. No superlatives about our company.
- End with three FAQs the READER would ask and one next step that
  FACTS shows we can deliver.

• [APPROVED OUTLINE]: the outline after your edits, not the model's first version.

The bracketed labels are scaffolding. They exist so that you can check each sentence against its source, and they come out before publishing. A sentence with a number and no label is the first thing to look at.

Prompt 4: List every checkable claim

Run this in a fresh conversation so the model reads the draft cold.

Do not rewrite anything. Read the draft below and list every
checkable claim in a table with these columns:

1. The exact sentence
2. Type: company fact / technical fact / number / comparison
3. Source label given in the draft, or "none"
4. What a person must confirm, and who (engineer, sales, owner)

Include claims in the title, tables, FAQ and call to action.
Then list separately: any sentence that implies a certification,
test, customer or result; and any example that could be mistaken
for a real one.

DRAFT: [DRAFT]

This gives you a worklist, and you still have to do the checking. The model can miss claims, and it cannot know whether your factory holds a tolerance. Open each source yourself; send the company facts to their owner. The AI article fact-check guide walks through that review.

Prompt 5: Edit for clarity without touching facts

Edit the draft below for clarity and length only.

- Do not change any number, unit, condition, product name or URL.
- Do not add any new fact, example or claim.
- Do not remove qualifiers such as "typically", "in most cases",
  "for this grade".
- Cut repeated points and background the READER already knows.
- Break paragraphs longer than five sentences.

After the edited text, list every sentence where you changed more
than word order, so I can recheck them.

READER: [READER]
DRAFT: [CHECKED DRAFT]

The closing list matters. Edits for flow are where a "typically" disappears, and it lets you reread only what moved.

Prompt 6: Title and meta description

Write 5 title options and 3 meta descriptions for the article below.

Titles: under 60 characters, state the question or the answer,
no "ultimate", "complete" or "guide to everything".
Descriptions: under 155 characters, say what the reader will be
able to decide after reading. Promise nothing the article does not
contain.

MAIN QUESTION: [ONE QUESTION]
ARTICLE: [FINAL TEXT]

The character limits are working conventions, not Google rules; search results truncate text to fit the device. Google writes its own search snippets mainly from the page content and uses the meta description when that describes the page better, so an accurate description is worth more than a clever one.

Example: Prompt 2 filled in

Example (fictional supplier). Marlow Fixings is an invented stainless fastener supplier. Its marketer has interviewed the applications engineer and fills in the outline prompt like this:

READER: Site engineer specifying fixings for outdoor steelwork
within sight of the sea. Knows fasteners, not metallurgy.

MAIN QUESTION: Should I specify A2 (304) or A4 (316) fasteners
for a coastal installation?

SCOPE: Bolts, nuts, washers M6 to M24. Not structural design.
Not duplex or coated carbon steel.

FACTS:
- We stock A2 and A4 in M6 to M24; other sizes to order.
- We supply a material certificate on request with each batch.
- Our engineer's starting point: consider A4 for salt exposure,
  then check temperature, exposure, crevices and drainage before
  recommending it. Refer immersed or stagnant seawater service
  for specialist review. Consider A2 for dry indoor or inland use.
- We do not advise on structural loading.

SOURCES:
- SOURCES-1: excerpt from our own product page on grade markings.
- SOURCES-2: engineer's email of 3 October answering my questions.
- SOURCES-3: BSSA guidance on 316 in seawater applications:
  https://bssa.org.uk/bssa_articles/selection-of-316-304-and-303-types-of-stainless-steels-for-seawater-applications/

LINKS: /products/stainless-bolts, /quality

OPEN:
- Delivery time for non-stock sizes: not confirmed.
- Whether we can supply to a specific project standard: ask sales.

The technical source matters here. BSSA's guidance on 316 in seawater explains that 316 can suffer pitting and crevice corrosion; temperature, exposure and drainage affect its suitability. Assess those conditions before making a final A4/316 recommendation for a coastal installation.

Look at what this packet contains: one question, a conditional recommendation, a boundary for specialist review, and two unknowns listed as unknowns. The outline should explain when A2 may be considered and which conditions still need checking. It will also flag delivery time as missing where a looser prompt would have invented "2 to 3 weeks".

What prompts cannot do

No wording makes a model know your factory. If FACTS is empty, the best prompt on this page produces a tidy generic article. Any sentence about your capabilities that you did not supply should be treated as invented until someone confirms it.

Assigning a role ("act as a senior materials engineer") changes the vocabulary and leaves the knowledge where it was. It can make invented detail sound more authoritative, which is worse.

And a prompt is a request the model can ignore. It may skip a rule, especially in a long draft. That is why Prompts 4 and 5 exist: the checking has to happen outside the instruction.

If you would rather not manage prompts at all, Blogo has no prompt box. It gathers the facts by asking the owner three to five questions before writing, and a figure appears in the draft only if it is in those answers or on the page linked in that sentence. How Blogo writes an article describes the stages.

FAQ

Can one prompt write the whole article?

It can write a whole article. You then have to find the invented parts yourself, with nothing to guide you. Six short steps take longer to run and much less time to check.

Do these prompts work in ChatGPT, Claude and Gemini?

Yes. They rely on plain instructions and labelled inputs and use no feature specific to one product. Output differs between models, so run Prompt 4 whichever you use.

Should I paste in competitor articles as examples?

No. The model will borrow their structure and sometimes their numbers. Paste your own best page if you want a style reference.

How do I stop the AI inventing statistics?

Supply the numbers you have, list the ones you lack under OPEN, and instruct it to write a marker where one is missing. Then check: Prompt 4 lists every number with its source so the unsourced ones stand out.

Can the model check its own citations?

It can list them. It cannot be relied on to confirm that a source says what the sentence claims. Open each one.

Where should we keep the prompts?

With the evidence packet and the approved outline for each article, so that anyone can see later where a statement came from. When the draft is ready, the WordPress AI writing workflow covers getting it onto the site as a draft.